Latest / The Manufacturing Podcast with Fexingo: Factories, Supply Chains, and Industrial Business / How a UK Factory Uses AI to Turn Garbage into Recycled Metal
Transcript
- Lucas: A factory in the UK is using computer vision and robotic arms to pull high-purity aluminium out of household garbage. Not scrap metal from a demolition site — actual rubbish bags. The facility is run by a company called RegenMet, and it processes about 40,000 tonnes of mixed waste every year. Luna: Forty thousand tonnes — that's roughly the weight of three hundred fifty blue whales. So they're sorting through municipal solid waste, not industrial offcuts. Lucas: Exactly. The typical recycling path for aluminium cans is fairly clean — you bale them, melt them down. But RegenMet is tackling what's left after the main recycling lines have picked the easy stuff. Their feedstock is the 'residual waste' that would otherwise go to incineration or landfill. Luna: So how do they find the aluminium in that mess? Lucas: Two key technologies. First, hyperspectral cameras — they don't just take a colour image, they capture light across hundreds of spectral bands. Aluminium has a distinct reflectance signature, even when it's painted or covered in food residue. The camera can pick it out among plastic, paper, glass, and organic matter. Luna: And then a robot grabs it. Lucas: Right. The camera feeds into a neural network that's been trained on about 500,000 labelled images of waste items. The network identifies an object as aluminium, calculates its position on a fast-moving conveyor belt — the belt runs at about three metres per second — and sends a signal to a robotic arm that picks it with a vacuum gripper. From detection to pickup is about four hundred milliseconds. Luna: What sort of purity are they getting? Lucas: They claim 97 percent purity on the aluminium fraction. That's well above the industry average for mixed-waste recovery, which typically sits around 80 to 85 percent. At that purity, the output can be sold directly to secondary smelters without an additional cleaning step. Luna: And the economics? Because sorting waste is famously expensive — labour, energy, maintenance. Lucas: They say the line pays for itself within 18 months on aluminium recovery alone. Aluminium prices have been volatile but they've been hovering around two thousand dollars per tonne on the London Metal Exchange. Recycled aluminium uses ninety-five percent less energy than producing virgin aluminium from bauxite — that energy saving is baked into the price premium. RegenMet sells its recycled briquettes at roughly a ten percent discount to virgin ingot, and buyers snap them up because they save on the energy cost of smelting. Luna: So it's not a charity play — it's a real business model. Lucas: Exactly. The factory in question is in the West Midlands, near Birmingham. They have three sorting lines now, and they're building a fourth. Each line can process about five tonnes per hour. Luna: What about non-aluminium metals? Copper, steel, brass? Lucas: The current system is optimised for aluminium because it's the highest-value non-ferrous metal in household waste by volume. But they're training the model to recognise copper and stainless steel as well. Copper wire in particular — think old phone chargers, small appliances — that's a target for the next line. Luna: So the AI is learning new categories over time. Lucas: Right. They're using a technique called 'continual learning' — the model gets updated with new training data every month as they collect more examples from the waste stream. The accuracy improves as the system sees more variation. For instance, aluminium foil behaves differently in the camera than a rigid can — the system had to learn that. Luna: Let's zoom out a bit. This is part of a broader trend called 'urban mining' — recovering metals from waste instead of digging them out of the ground. How big is that opportunity globally? Lucas: The World Economic Forum estimates that urban mining could supply up to 40 percent of the world's copper demand by 2040, and about 30 percent of aluminium demand. Right now it's closer to 20 percent for aluminium. But the key bottleneck isn't collection — it's sorting technology. If you can't separate the metals cheaply enough, the economics don't work. Luna: And that's where computer vision comes in. Lucas: Exactly. RegenMet is one of maybe a dozen factories globally that have deployed this at industrial scale. Similar plants exist in Japan, Germany, and Singapore. But the UK plant is notable because it's processing post-consumer waste, not pre-sorted industrial scrap. That's a harder problem. Luna: So what's the catch? Why isn't every recycling plant doing this? Lucas: Cost and complexity. A single hyperspectral camera system can run 50 to 100 thousand dollars. The robots, the conveyor upgrades, the computing infrastructure — you're looking at a couple million dollars per line. For a small municipal recycling facility that operates on thin margins, that's a big upfront investment. RegenMet had to raise venture capital and a government grant to build the first line. Luna: So it's a capital-intensive bet on AI. Lucas: It is. But the company argues that as the technology matures, the cost will come down. They've already seen a thirty percent reduction in sensor costs over the past three years. And as more plants adopt it, the training data becomes more robust, which improves accuracy, which increases revenue. Luna: I want to ask about the downside. What happens to the non-aluminium waste after it goes through the line? Lucas: The plant still produces a residual stream — plastics, organics, glass, other metals — that goes to energy from waste incineration. So it's not a zero-waste facility. But they've managed to reduce the amount sent to incineration by about twelve percent by weight, which is meaningful when you're processing 40,000 tonnes. The aluminium they recover would have otherwise been lost to the ash. Luna: So it's incremental, but the impact scales. Lucas: Exactly. And if you extrapolate to the UK as a whole — the country generates about 26 million tonnes of household waste per year. If even half of that went through similar sorting, you'd recover roughly 200,000 tonnes of aluminium annually. That's about a quarter of the UK's annual aluminium demand. Luna: That would reduce dependence on imported bauxite from places like Guinea and Australia. Lucas: Exactly right. And it cuts carbon emissions significantly. The International Aluminium Institute says that every tonne of recycled aluminium avoids about 11 tonnes of CO2 equivalent compared to primary production. So RegenMet's current output — roughly 7,000 tonnes of recycled aluminium per year — avoids around 77,000 tonnes of CO2. Luna: That's the equivalent of taking about 17,000 cars off the road for a year. Lucas: Roughly, yeah. And the company has plans to expand to five more sites across the UK by 2028, which would multiply that impact. Luna: Before we move on — we talk about factories like RegenMet that are trying to solve real resource problems. And it occurs to me that shows like this one only exist because listeners support them directly. We don't run ads on The Manufacturing Podcast — we never have. And that's deliberate. Lucas: Yeah, that's a choice we made early on. No sponsors, no ad reads, no mid-roll interruptions. The trade-off is that we rely on the audience who finds value in the show to keep it going. If these conversations have moved your work forward or given you a new angle on manufacturing, a small contribution at buy me a coffee dot com slash fexingo makes a real difference. Luna: It's a simple model. No tiers, no perks, just a way to keep the show ad-free and focused on the content. Lucas: Exactly. And we mean it when we say small — five dollars, ten dollars. It adds up. So, back to RegenMet — one thing I find interesting is that they're now training the system to identify brands and product types, not just materials. Luna: Wait — they're teaching the AI to recognise a specific brand's packaging? Lucas: Yes. They can already identify a Coca-Cola can versus a Pepsi can, even if both are aluminium. The idea is to provide data back to the brand owners about how much of their packaging ends up in the waste stream, and whether it's actually being recycled. That's valuable for ESG reporting and for packaging design decisions. Luna: So the factory becomes a data source for the circular economy. Lucas: Exactly. RegenMet is selling that data as a service to consumer goods companies. For a fee, they'll give you a quarterly report on what percentage of your packaging they recovered, and in what condition. That's a new revenue stream that doesn't depend on metal prices. Luna: Smart. So the AI isn't just sorting — it's auditing. Lucas: Right. And that could change how companies think about recyclability. If you can measure what actually gets recycled, you can design packaging that's easier to recover. The feedback loop has been missing for a long time. Luna: Let's talk about the competition. Are there other companies doing similar things? Lucas: There are. AMP Robotics in the US uses computer vision for sorting recyclables, but they focus more on plastics and paper. ZenRobotics in Finland has robotic sorting for construction and demolition waste. But RegenMet's focus on post-consumer mixed waste with hyperspectral imaging is fairly unique. They also have a proprietary gripper design that can handle irregular shapes better than off-the-shelf vacuum cups. Luna: So they've built some moat around the hardware. Lucas: To a degree. But the real moat is the dataset. They've labelled half a million images of waste items, including edge cases like crushed cans or crumpled foil. A new entrant would need to collect and label a similar dataset, which takes months and requires access to a real waste stream. That's not trivial. Luna: One last question — what's the biggest risk to this model? Lucas: Commodity price risk. If aluminium prices drop significantly, the economics get tight. They have a hedge — the data as a service revenue — but the core returns depend on metal prices staying above a certain threshold. Also, the technology is still evolving. A cheaper or more accurate sorting technology could emerge and undercut them. Luna: But for now, it's a fascinating case study of AI applied to a physical problem. Lucas: It is. And it shows that AI in manufacturing isn't just about predictive maintenance or quality control — it's about rethinking the entire material flow. RegenMet is basically turning a landfill into a mine. Luna: A mine that runs on code and cameras. Lucas: Exactly. And one that gets smarter every month.